MXD Process Embeds AI in Workflows to Cut Quoting From Days to Minutes

Hard-Fought Lessons from a Year of Scaling AI in an SMB With MXD Process — Ai4

Ai4

Standalone AI chatbots fail in small businesses unless models are embedded directly into native operational workflows guided by domain experts and clean data plumbing.

AI-assisted summary of Hard-Fought Lessons from a Year of Scaling AI in an SMB With MXD Process — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 15, 2026, this presentation features Eric Seiberling, Vice President of Sales and Marketing at MXD Process. According to the description, Seiberling shares how a small industrial manufacturing company scaled AI across sales, support, and two acquisitions without consultants or an IT team, delivering a practical playbook on what to automate, where humans matter, and how to budget effectively.

## TL;DR Eric Seiberling recaps the second year of implementing AI at MXD Process, a $30 million industrial equipment manufacturer, moving beyond initial custom GPT wins to address real-world operational bottlenecks. He explains how standalone chat interfaces failed the majority of employees through prompt friction, uncritical acceptance of hallucinations, and manual 'digital coal' data wrangling. To overcome this, the company embedded AI invisibly into existing workflows, rapidly vibe-coding support dashboards, spinning up CRM instances, and building a 24/7 custom tank quoter with systems integrator Level 12. The resulting playbook emphasizes that backend data quality dictates AI performance, token costs are negligible compared to leadership attention, and organizations must actively train and enforce critical human judgment.

## Atmosphere & style The talk is refreshingly pragmatic, energetic, and candid, eschewing tech-industry buzzwords in favor of battle-tested operational honesty. Seiberling engages the audience with self-effacing humor—cracking jokes about freezing slide decks, soccer stoppage time, and his own tendencies to fall into 'duct-tape traps' while manually shoveling data. It feels like an authentic dispatch from the industrial trenches, contrasting Silicon Valley vendor hype with down-to-earth lessons that empower non-technical operators to trust their own practical common sense.

## Key takeaways - [01:29](https://www.youtube.com/watch?v=jAO7jheIl0k&t=89s) Initial custom GPT deployments delivered dramatic early gains for MXD Process, reducing lead response time from 5 days to 1 day and raising deal close rates from 17% to 27%. - [03:25](https://www.youtube.com/watch?v=jAO7jheIl0k&t=205s) Standalone chatbot interfaces create severe friction and fail for the majority of non-technical workers, leading to employee disillusionment, legacy pushback, or blind over-reliance on hallucinations. - [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) The risk of 'confident failure' occurs when articulate, professionally formatted AI outputs mask omitted data or flawed conclusions, allowing bad insights to circulate unchecked. - [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) Upgrading to smarter language models cannot overcome dirty data pipelines; an organization's generative capabilities are strictly capped by its backend data plumbing. - [08:01](https://www.youtube.com/watch?v=jAO7jheIl0k&t=481s) To achieve lasting adoption, AI must be extracted from isolated browser windows and embedded invisibly into the tools and daily workflows where employees already work. - [10:51](https://www.youtube.com/watch?v=jAO7jheIl0k&t=651s) AI can rapidly accelerate administrative setup, demonstrated by using Claude Co-work to configure an entire fresh HubSpot CRM instance in four hours from raw sales process notes. - [15:20](https://www.youtube.com/watch?v=jAO7jheIl0k&t=920s) Vibe coding rapid prototypes must be paired with experienced systems integrators to ensure transactional scalability and avoid the failure modes of purely generated bespoke software. - [17:09](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1029s) For small and midsize businesses, software licenses and API token expenses are negligible rounding errors; the true scarce resource that requires strict budgeting is leadership attention and experimentation bandwidth. - [19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) AI compresses domain experts' execution time but cannot manufacture expertise from scratch, making direct involvement from experienced business leaders essential during builds. - [19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) Organizations must actively train, evaluate, and manage staff on critical thinking and judgment to prevent uncritical reliance on generative outputs. - [20:57](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1257s) AI adoption truly succeeds when the technology stops looking like a standalone novelty and operates invisibly in the background of everyday business processes.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=jAO7jheIl0k&t=4s) Introduction and the Industrial Equipment Context

Eric Seiberling introduces himself as the Vice President of Sales and Marketing at MXD Process. Acknowledging that most people in the room likely haven't heard of the business, he describes MXD Process as a $30 million industrial equipment manufacturer. Rather than making tanks designed for combat, the company builds heavy-duty stainless steel tanks and mixing vats used to blend chemicals and food ingredients for everyday products. Their equipment spans a vast range of manufacturing needs, from blending sweet barbecue sauces—including well-known names like Sweet Baby Ray's and numerous private-label brands for major grocery stores—all the way to high-spec aerospace coatings.

When Seiberling joined the company three years ago, the business was in need of an operational overhaul. Tasked with fixing core commercial processes, he turned to AI. Having presented their initial exploratory stage at the previous year's Ai4 conference, he returns to deliver the sequel to that story. Mid-presentation, Seiberling's slide deck hits a technical snag and freezes, prompting him to joke about whether he gets extra stoppage time like in a soccer match before pressing forward without missing a beat.

- **Company profile:** MXD Process, ~$30M industrial equipment manufacturer specializing in stainless steel mixing tanks. - **Product applications:** Private-label and brand-name food processing (e.g., Sweet Baby Ray's barbecue sauce) up to technical aerospace coatings.

---

[01:29](https://www.youtube.com/watch?v=jAO7jheIl0k&t=89s) The Early Wins: Crawl, Walk, Run and Custom GPTs

Recapping the initial phase shared in his prior talk, Seiberling explains their "crawl, walk, run" deployment strategy, starting strictly with tools the company already possessed. They focused on building automated workflows in HubSpot and creating custom GPTs. General off-the-shelf models like GPT-3 were utterly incapable of handling industrial chemical engineering; Seiberling compares asking GPT-3 to solve their engineering problems to "having a fifth grader try to do chemistry."

To overcome this, MXD Process engineered specialized custom GPTs containing domain-specific guardrails and chemical formulas. The early quantitative payoff was dramatic across three key business metrics:

These dramatic gains were transformative for a small business. Following the initial talk, Seiberling was mobbed with inquiries and felt confident they had solved the transition from zero to one. Leaving the conference, he patted himself on the back, convinced they possessed the definitive playbook to scale. But as year two unfolded, he was met with a stark reality check: "I was wrong." What felt like crossing the finish line of the Boston Marathon turned out to be the starting line of an American Ninja Warrior obstacle course.

---

[03:25](https://www.youtube.com/watch?v=jAO7jheIl0k&t=205s) The Container Problem: Why Chatbots Fail the Majority of Users

The primary bottleneck MXD Process encountered during rollout was that custom GPTs were trapped in the wrong software container. Relying on a chat interface introduces severe user-friction hurdles: an employee must open a separate browser tab, log into a service, remember to frame and refine complex prompts, wait for output, and copy-paste it elsewhere. This architecture demands an inherently proactive, motivated user. While about a third of the MXD Process team took to large language models immediately because they already understood and trusted them, the remaining two-thirds faltered.

```mermaid flowchart TD accTitle: Custom GPT User Segmentation A["Rollout of Custom GPTs ('Arlin')"] --> B["Motivated Champions (~33%) Tech-fluent, prompt well, gain real value"] A --> C["Friction & Disillusionment Struggled with prompt phrasing; quit after bad output"] A --> D["Traditional Veterans 30+ years selling via phone; ignored the tool completely"] A --> E["Uncritical Over-Believers Treated AI output as gospel; pitched unsellable solutions"] ```

Seiberling found that the unmotivated or inexperienced workforce broke down into three failure groups: 1. **The Disillusioned:** Employees who attended training, attempted basic prompts, received lackluster answers, declared "this sucks," and abandoned the tool entirely. 2. **The Traditionalists:** Veteran industrial sales reps who had sold tanks over the phone for 30 years and refused to let "some computer" instruct them on equipment sales, effectively ghosting the tool. 3. **The Uncritical Believers:** Staff who consulted their chemical engineering knowledge assistant—internally christened "Arlin"—and treated its generative outputs as absolute truth. When dealing with intricate chemical specifications, Arlin occasionally hallucinated or suggested bizarre, convoluted configurations designed solely to force-fit MXD Process equipment. Employees took these outputs directly to prospective clients, forcing Seiberling to intervene: "Wait, wait, wait, wait, wait, we don't sell that. Stop."

---

[05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) Confident Failure and the "Digital Coal" Trap

The risk that rattled Seiberling the most was what he terms "confident failure." A newly hired business manager used Claude to parse commercial performance data and generated an elegant, beautifully formatted business report. However, the underlying analysis omitted critical datasets and produced fundamentally incorrect conclusions. In traditional workflows, a hesitant junior employee would normally ask, *"Hey Eric, could you check my work?"* But because the AI generated an articulate, authoritative-sounding briefing, the manager distributed it company-wide without validation, leading corporate leadership to begin operating on flawed data.

```mermaid flowchart LR accTitle: The Digital Coal Human Integration Bottleneck subgraph Sources ["Siloed Raw Data"] ERP["ERP System"] CRM["CRM (HubSpot)"] Ship["Freight Broker Data"] Reps["Ad-hoc Reports"] end subgraph Bottleneck ["The Human Plumbing Layer"] Me["Eric Seiberling ('Shoving Digital Coal') Manual Extraction & Cleansing"] end subgraph UI ["Downstream Artifacts"] Vibe["Vibe-Coded Executive Dashboard"] end ERP --> Me CRM --> Me Ship --> Me Reps --> Me Me --> Vibe ```

Attempting to resolve data presentation issues, Seiberling "vibe coded" a custom executive dashboard. While visually stunning and insightful for decision-making, it generated an accidental operational trap. Rather than automating his workload, Seiberling found himself acting as the literal integration layer: - Disparate records had to be stitched together manually across the company's enterprise resource planning (ERP) platform, CRM, external shipping broker logs, and disparate spreadsheets. - Every single Monday morning, he had to extract the raw metrics, cleanse them, reconcile errors, and manually upload them into the dashboard. - As Seiberling put it, he was the guy "shoving digital coal into the AI and into the dashboard."

The core takeaway was unmistakable: "Nobody's ever fixed dirty data by upgrading to a better model." Advanced intelligence cannot overcome poor underlying data pipelines, and a business's generative ceiling is always dictated by its backend plumbing.

---

[08:01](https://www.youtube.com/watch?v=jAO7jheIl0k&t=481s) The Duct-Tape Trap and Embedding AI into Native Workflows

Shortly after experiencing these bottlenecks, MXD Process acquired an environmental services company. Pressured by a lack of time and urgently needing to equip the acquired unit, Seiberling fell into the exact same trap he had just identified. Despite knowing the architectural flaws of standalone custom GPTs, he defaulted to the familiar "duct tape" workaround because crisis scenarios force managers to reach for immediate quick fixes rather than sustainable plumbing.

To truly make AI stick in small-to-medium businesses, organizations must extract AI out of isolated browser windows and embed it directly into the natural flow of daily work. Driving adoption cannot hinge on urging employees to adopt new habits or learn prompting techniques. Instead, workflows must be systematically redesigned so that models operate invisibly in the background, surfacing insights where team members already carry out their jobs.

Seiberling cautions that commercial software vendors routinely make sweeping promises, yet standard solutions fail in long-lead, capital-intensive manufacturing environments where every sale demands bespoke engineering and complex chemistry.

Illustrating how fast embedded discovery can work, Seiberling discusses his role as fractional Chief Revenue Officer (CRO) for Albert Paper Products, a manufacturer of tertiary packaging materials (such as the U-boards holding chocolate bars in candy wrappers or custom die-cut cosmetic inserts). Albert Paper was still managing its entire commercial pipeline using manual Excel spreadsheets and physical paper.

Rather than forcing complex software upon them, Seiberling sat down with their leadership for two full days with meeting recorders active (such as Otter, Granola, or Fireflies): - Every conversation, historical nuance, and operational hurdle was captured verbatim. - Using generative tools to distill the transcripts, Seiberling constructed a modern commercial sales architecture in under 24 hours. - The resulting overhaul delivered customer personas, Ideal Customer Profiles (ICPs), and competitive positioning matrices—instantly giving a paper-bound manufacturer modern sales capabilities without forcing sales reps to interact with raw chat windows.

[10:51](https://www.youtube.com/watch?v=jAO7jheIl0k&t=651s) Fast-Tracking HubSpot Setup and Support Analytics via Claude

Armed with notes from a three-day intensive workshop with the sales team, the next challenge was standing up HubSpot from scratch for a newly acquired entity. Rather than enduring weeks of tedious manual field mapping and pipeline configuration, the entire setup was accelerated using Claude Co-work on a newly purchased Mac Studio. By feeding the AI three core inputs—an Excel sheet of contacts, a spreadsheet of companies, and the newly documented sales process—along with the prompt *"This is a blank HubSpot instance. Please stand it up,"* the model only needed to ask four or five clarifying questions. In roughly four hours, the instance was fully configured and running according to the exact specifications of the sales experts, eliminating weeks of administrative grunt work.

```mermaid flowchart TD accTitle: Rapid HubSpot Configuration and Closed-Loop Support Workflow subgraph Inputs["1. Business Inputs & Experts"] A["Sales Process Document"] B["Contacts & Companies Excel"] C["Domain Experts in the Room"] end

subgraph AI_Engine["2. AI Processing"] D["Claude Co-work on Mac Studio"] E["Automated Configuration & Analysis"] end

subgraph Outcomes["3. Deployed Systems"] F["Configured HubSpot Instance (4 Hours Setup)"] G["Support Ticket Interventions & Preventative Content Engine"] end

Inputs --> AI_Engine AI_Engine --> Outcomes ```

Applying the same rapid approach to customer service, an operational support dashboard was vibe coded in under an hour against HubSpot support data. Working directly alongside the support lead, the system analyzed historical ticket trends to identify critical operational interventions. These insights were fed back into Claude to generate targeted help content that prevents recurring customer issues at the source, establishing a closed-loop system. Because the underlying data plumbing was already connected, daily data cleanup and migration routines taking two hours were reduced to hitting refresh, paving the way for automated agent-driven internal distribution.

- **HubSpot deployment timeline:** 4 hours total to full operational readiness versus multiple weeks of manual setup. - **Clarification efficiency:** Claude required only 4–5 targeted questions to map out custom fields, objects, and pipeline stages. - **Support analytics vibe coding:** Complete customer support dashboard built in less than 60 minutes. - **Maintenance reduction:** Eliminated approximately 2 hours per day previously spent manually cleaning, transforming, and migrating data.

---

[13:11](https://www.youtube.com/watch?v=jAO7jheIl0k&t=791s) Vibe Coding the Heritage Equipment Tank Quoter

Following the acquisition of Heritage Equipment Company, quoting custom stainless steel processing tanks remained a major operational bottleneck, routinely taking five to six business days per request. Operations had spent three years unsuccessfully trying to produce a workable spreadsheet quoting matrix. To solve this, AI was utilized not as an end-user black box, but as a rules engine. The team leveraged Arlin—a specialized chemical engineering GPT—to establish the intricate physical and safety constraints required for manufacturing (such as sanitary design standards for food products like barbecue sauce versus explosion-proof engineering for caustic chemicals).

```mermaid flowchart LR accTitle: Heritage Tank Quoter Architecture and Development Flow subgraph Engineering["1. Domain Logic"] A["Arlin Chemical GPT (Rules & Constraints)"] end

subgraph Development["2. Co-Development Sprint"] B["Frontend Vibe Coded in Claude Co-work"] C["Level 12 Integration (Robust Backend Architecture)"] end

subgraph Deployment["3. Live Quoting Portal"] D["Instant Customer Quotes (Available 24/7 in Minutes)"] end

Engineering --> B B <--> C C --> D ```

The frontend application was vibe coded in Claude Co-work while Bruce, the owner of boutique systems integrator Level 12, engineered the backend architecture in parallel. This hybrid workflow produced a fully functioning, scalable prototype in under three weeks, with five additional weeks dedicated to data ingestion and rigorous model validation. Launched just nine days prior to the presentation, the tool transformed a multi-day turnaround into a self-service web application where customers can configure batch sizes and receive instant quotes at 11:00 PM. The project generated significant industry attention, including coverage on Yahoo Finance, proving that even legacy manufacturing sectors can deploy modern software experiences.

- **Quote turnaround compression:** Reduced quoting time from 5–6 business days to immediate, on-demand calculation. - **Rapid co-development:** Working full-stack prototype completed in under 3 weeks; total project launch in 8 weeks (including 5 weeks of data validation). - **Engineering compliance:** Arlin codified complex constraints across sanitary food-grade specifications and hazardous chemical ratings. - **Audience trial:** Attendees testing the live quoting tool via QR code were advised to append "AI4" to their name to prevent automated sales outreach.

---

[15:20](https://www.youtube.com/watch?v=jAO7jheIl0k&t=920s) Why Systems Integrators and Domain Experts Beat DIY SaaS

The success of these technical builds stems from an unapologetically business-first orientation rather than traditional IT bureaucracy. Without an internal software development staff—relying on only a single network administrator—domain knowledge had to lead technical implementation. Having business experts actively participating in the room during builds delivers two decisive benefits: the software is architecturally accurate from day one, and internal user adoption is immediate because the team co-authored the solution. This collaborative reality directly counters the traditional enterprise IT trap of disappearing into a closet with requirements and returning months later with an unusable product.

Crucially, vibe coding alone is insufficient for mission-critical enterprise systems. Partnering with Level 12 provided the software engineering guardrails necessary to prevent UX and architectural failures. Those preaching the "SaaS apocalypse" and claiming anyone can vibe code a bespoke CRM fail to recognize that purely generated software collapses under real-world transactional weight and scale. Pairing domain experts with experienced systems integrators ensures that AI-accelerated prototypes evolve into resilient, production-grade applications.

- **Team composition:** Led by commercial leadership with zero internal software developers and one network infrastructure manager. - **Co-creation over requirement handoffs:** Direct domain expert collaboration avoids the cycle of isolated development and post-launch rejection. - **SaaS apocalypse reality check:** Disposable vibe-coded tools fail under enterprise load; professional backend integration remains essential for longevity.

---

[17:09](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1029s) SMB Tokenomics: Budgeting Attention Over Software Licenses

Tokenomics and API licensing costs present fundamentally different trade-offs for small-to-midsize businesses compared to corporate giants like Amazon or Procter & Gamble. While large enterprises must police massive token consumption and recurring seat costs, SMB software and token expenses are effectively *"rounding errors"* relative to their return on investment. The actual scarce resource requiring strict budgeting is organizational time and leadership attention.

Because progress depends entirely on focused momentum, an AI initiative stalls completely the moment leadership or frontline subject matter experts become *"too busy"* to engage. Consequently, capital is far better spent funding rigorous data cleanup and hiring technical integration partners than obsessing over license fees. Without clean underlying data plumbing, AI implementations will fail entirely or lead to expensive, flawed business decisions.

---

[19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) The Five Playbook Lessons for Scaling SMB AI

Reflecting on a full year of scaling AI across core operations, five hard-earned principles define the operating playbook for small and medium businesses:

```mermaid flowchart TD accTitle: The Five Playbook Lessons for Scaling AI in SMBs L1["1. Get It Out of the Chat Window (Embed directly into workflows)"] L2["2. Get an Expert in the Room (Compress time, don't fake expertise)"] L3["3. Your Ceiling Is Your Data (Fix foundational plumbing first)"] L4["4. Budget Attention, Not Licenses (Focus on bandwidth, cut unused seats)"] L5["5. Teach and Enforce Judgment (Hire, evaluate, and fire on critical thinking)"]

L1 --> L2 --> L3 --> L4 --> L5 ```

1. **Get it out of the chat window:** When AI is confined to standalone browser chat interfaces, employees abandon it during busy periods and revert to old habits. Tools must be hardwired into existing daily workflows. 2. **Get an expert in the room:** "AI compresses experts' time, but it doesn't manufacture expertise." Without domain knowledge guiding the model, output quality degrades rapidly. 3. **Your ceiling is your data:** Faulty data plumbing guarantees AI failures or costly miscalculations. Clean data must precede advanced automation. 4. **Budget attention, not licenses:** License optimization matters far less than preserving cognitive bandwidth for experimentation. Inactive seats (such as ten recently pruned ChatGPT licenses) can simply be cut without affecting output when AI is baked directly into the process. 5. **Teach and enforce judgment:** Guarding against cognitive decay requires cultivating critical thinking across the organization. Employees must be hired, evaluated, and if necessary fired based on their curiosity and willingness to rigorously question AI-generated outputs rather than accepting answers blindly.

[20:57](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1257s) Making AI Invisible: Why Adoption Succeeds When AI Stops Looking Like AI

Eric Seiberling distills his fundamental philosophy on artificial intelligence adoption down to a single core truth: "AI only sticks when it stops looking like AI." Reflecting on earlier tech waves, he notes that he lived through the birth of the consumer internet and the dot-com boom, drawing a direct parallel to the rise of e-commerce. During his tenure at Procter & Gamble, industry prognosticators routinely sounded the alarm that traditional retail was doomed to die. Yet retail never died; it simply evolved. Today, online shopping does not feel like a disruptive technical novelty called "e-commerce"—it has simply become the ordinary way people shop.

In the exact same way, AI will only achieve lasting success in business operations when companies stop treating it as a novel destination or an isolated novelty. True victory comes when organizations pull AI out of standalone chatbot windows, weave it seamlessly into existing operational workflows, and render it completely invisible to everyday users so that it operates frictionlessly in the background. As Seiberling describes it, the ultimate goal is making operations genuinely "automatic."

```mermaid flowchart LR accTitle: Evolution from Standalone Tool to Invisible Process subgraph S1["Novelty Stage"] A["Standalone Chatbot"] --> B["Separate AI Destination"] end subgraph S2["Integrated Stage"] C["Embedded in Workflows"] --> D["Invisible to End User"] end B -.->|"Evolution"| C ```

- **Historical Parallel**: Recalls the doom-and-gloom predictions at Procter & Gamble claiming e-commerce would kill brick-and-mortar retail, demonstrating that transformative tech simply integrates into everyday life rather than destroying existing models. - **The True Benchmark of AI Success**: Moving away from user-facing prompts and chatbots toward deeply embedded, invisible system processes that execute automatically.

---

[21:44](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1304s) Parting Advice: Trusting Your Judgment and Connecting

Closing out his presentation, Seiberling looks ahead to returning the following year with another update on the company's continuous operational journey. He invites audience members to connect with him on LinkedIn, though not without laying down a few humorous ground rules: non-competitors are fully welcome to reach out and swap notes; direct competitors are welcome to talk too, but he jokes that he will intentionally feed them "all the wrong information"; and eager software vendors will be granted a strict, non-negotiable thirty-second pitch window.

His final takeaway for leaders trying to implement AI inside small and midsize businesses is grounded in confidence over hype. Rather than feeling intimidated by high-cost enterprise consultancy narratives or complex toolchains, small operators should back their own instincts: "Trust yourself. Trust your judgment. Take these lessons and you'll be able to do the same things that we've been able to do."

- **Rules of Engagement for Connecting**: - **Peers and Non-Competitors**: Open invitation to connect and discuss implementation strategies on LinkedIn. - **Competitors**: Still welcome to chat, with the humorous caveat that he will feed them misleading advice. - **Vendors**: Limited strictly to a 30-second elevator pitch. - **Final Charge**: Urges operators to rely on their practical business common sense rather than external hype when adopting artificial intelligence.

## Notable quotes > [01:29](https://www.youtube.com/watch?v=jAO7jheIl0k&t=89s) having a fifth grader try to do chemistry. — Eric Seiberling

> [01:29](https://www.youtube.com/watch?v=jAO7jheIl0k&t=89s) I was wrong. — Eric Seiberling

> [03:25](https://www.youtube.com/watch?v=jAO7jheIl0k&t=205s) Wait, wait, wait, wait, wait, we don't sell that. Stop. — Eric Seiberling

> [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) Hey Eric, could you check my work? — Junior Employee

> [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) shoving digital coal into the AI and into the dashboard. — Eric Seiberling

> [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) Nobody's ever fixed dirty data by upgrading to a better model. — Eric Seiberling

> [10:51](https://www.youtube.com/watch?v=jAO7jheIl0k&t=651s) This is a blank HubSpot instance. Please stand it up, — Eric Seiberling

> [19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) AI compresses experts' time, but it doesn't manufacture expertise. — Eric Seiberling

> [20:57](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1257s) AI only sticks when it stops looking like AI. — Eric Seiberling

> [21:44](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1304s) Trust yourself. Trust your judgment. Take these lessons and you'll be able to do the same things that we've been able to do. — Eric Seiberling

## Mentioned Ai4, Albert Paper Products, Amazon, Arlin, Claude, Claude Co-work, Eric Seiberling, Fireflies, Granola, Heritage Equipment Company, HubSpot, Level 12, Mac Studio, MXD Process, Otter, Procter & Gamble, Sweet Baby Ray's, Yahoo Finance

## Video description Eric Seiberling - VP of Sales and Marketing, MXD Process

Last year, I showed how we started using AI at a small industrial company—with no IT team and no fluff. Since then, we’ve scaled those early wins in sales, marketing, support across the company and two acquisitions — without consultants or custom software. But it wasn’t easy. In this talk, I’ll share what we broke, what we fixed, and the messy, behind-the-scenes lessons we learned trying to make AI stick in a real business. You’ll walk away with a practical playbook for growing AI inside a small team—what to automate, where humans still matter, and how to budget for the things no one warns you about.

[Source](https://www.youtube.com/watch?v=jAO7jheIl0k)

中文简报

MXD Process将AI嵌入工作流程，报价时间从数天缩短至数分钟在小型企业中，独立AI聊天机器人往往难以奏效，除非在领域专家和规范数据管道的指导下，将模型直接嵌入原生业务工作流中。 Eric Seiberling回顾了年营收3000万美元的工业设备制造商MXD Process落地AI第二年的历程，展示了公司如何超越最初自定义GPT带来的成效，转而解决现实中的业务运营瓶颈。他阐释了独立聊天界面为何未能惠及大多数员工：提示词门槛、对模型幻觉未经审视的盲目接受，以及手动“数字挖煤”式的数据整理。为克服这些问题，该公司将AI无缝嵌入现有工作流程，通过快速vibe-coding开发支持看板、搭建CRM实例，并与系统集成商Level 12合作构建了24/7全天候定制储罐报价工具。由此形成的实践指南强调：后端数据质量决定AI性能，相较于领导层的精力投入，Token成本微不足道，且企业必须主动培训并强化员工的批判性判断力。 MXD Process 通过早期部署定制 GPT 与自动化工作流，将销售线索首触响应时间从 5 天压缩至 1 天，报价速度提升 80%，交易成单率由 17% 提高至 27%。 (1:29) 独立的 Chatbot 交互界面存在严重使用摩擦，导致约三分之二的员工陷入幻灭、传统销售直接抵触，或因盲信 AI 幻觉而向客户推销不可行的方案。 (3:25) 企业大模型应用能力的上限严格受制于底层数据管线，更聪明的模型无法解决脏数据问题，且专业格式的输出容易引发掩盖事实缺陷的“自信失败”。 (5:45) 要实现真正的落地采纳，必须将 AI 隐形嵌入员工现有的日常工作流中；团队借助 Claude Co-work 仅用 4 小时便根据原始销售流程笔记搭建起完整的全新 HubSpot CRM。 (10:51) 对中小企业而言，软件订阅和 API Token 成本只是微不足道的“四舍五入误差”，最需要严格把控与预算的稀缺资源是管理层的注意力与试错带宽。 (17:09)

Reading summary

Scaling artificial intelligence within small and midsize businesses requires moving away from standalone chat interfaces and embedding models directly into existing operational workflows. Drawing from implementations at industrial equipment manufacturer MXD Process, Vice President of Sales and Marketing Eric Seiberling argues that foundational data pipelines, domain expertise, and critical human judgment matter far more than model selection or software licensing costs. Early Gains and the Chatbot Interface Bottleneck (1:29) MXD Process, an industrial manufacturer producing stainless steel mixing tanks for applications ranging from barbecue sauce to aerospace coatings, initially deployed a 'crawl, walk, run' automation strategy using HubSpot workflows and custom GPTs containing chemical engineering formulas. This early phase reduced first-contact lead times from 5 days to 1 day, accelerated quoting speed by 80%, and increased deal close rates from 17% to 27%. However, widespread rollout stalled because custom GPTs were confined to standalone chat windows, requiring employees to open separate tabs, formulate complex prompts, and transfer outputs manually. Seiberling found that roughly one-third of the workforce adopted the tools smoothly, while the remaining two-thirds struggled. These non-adopters split into three failure profiles: disillusioned staff who abandoned the tool after basic prompts yielded mediocre outputs; veteran sales reps with decades of phone-based selling who ignored the software; and uncritical believers who accepted generative outputs as fact. The third group caused commercial risk when an engineering assistant nicknamed 'Arlin' hallucinated equipment configurations, leading staff to pitch non-existent products to prospective clients. Data Pipeline Fragility and Confident Failure (5:45) Standalone generative models introduced what Seiberling describes as 'confident failure,' in which articulate, professional presentation masks severe analytical errors. In one instance, a business manager used Claude to summarize performance metrics into a polished executive briefing that omitted vital datasets and reached incorrect conclusions, which company leadership distributed internally without validation because it appeared authoritative. Efforts to solve data visibility by rapidly 'vibe coding' an executive dashboard created an operational bottleneck. Seiberling found himself acting as a human integration layer, spending every Monday morning manually extracting, cleansing, and reconciling siloed data from the company's enterprise resource planning platform, CRM, freight logs, and ad-hoc spreadsheets. Advanced intelligence cannot compensate for poor underlying data plumbing, reinforcing that an organization's AI capability is bounded by the quality of its backend data pipelines. Embedded Workflows and Hybrid Software Integration (8:01) To achieve lasting adoption, MXD Process transitioned from manual prompts to invisible background automations embedded within daily operations. Illustrating this approach at tertiary packaging firm Albert Paper Products, meeting recordings from executive interviews were parsed with language models to produce sales personas, ideal customer profiles, and competitive matrices in under 24 hours. Similarly, setting up a HubSpot instance from scratch for an acquired company—using Claude Co-work fed with company sheets, contact lists, and sales documentation—was completed in roughly four hours instead of weeks, requiring only four or five clarifying questions from the model. For complex commercial operations, MXD Process paired generative rapid prototyping with traditional software engineering. To address a legacy quoting process that took five to six business days at Heritage Equipment Company, the team vibe coded a customer-facing frontend prototype in Claude Co-work using Arlin's chemical and sanitary constraints, while systems integrator Level 12 developed the scalable backend architecture. The hybrid project delivered an instant self-service quoting application in eight weeks (including five weeks of data validation), illustrating that vibe coding alone cannot support production workloads without robust integration. SMB Economics: Allocating Attention Over Token Costs (17:09) Seiberling notes that small-to-midsize businesses face fundamentally different financial dynamics than large enterprises like Amazon or Procter & Gamble. For SMBs, software subscription and token expenses represent minor rounding errors compared to business return on investment. The primary bottleneck is scarce organizational attention, experimentation bandwidth, and the availability of frontline subject-matter experts. Because domain experts are required to establish accurate constraints and prevent software rejection, projects stall whenever key personnel lack the bandwidth to engage. SMB capital is therefore more effectively spent securing systems integration support and repairing foundational data hygiene rather than policing token usage. In MXD Process's experience, ten unused ChatGPT seats were pruned without hurting output once AI capabilities were directly embedded into automated backend routines. Seiberling concludes that AI adoption mirrors the evolution of e-commerce, which ceased being a disruptive novelty once it blended into standard retail. Realizing sustained value requires organizations to institutionalize critical thinking, hiring and reviewing employees based on their willingness to question AI outputs rather than accepting generative answers uncritically.

精炼阅读摘要

MXD Process销售与营销副总裁Eric Seiberling指出，中小企业规模化应用AI的真正瓶颈不在于基础模型的能力或软件许可费用，而在于脱离业务的工作流形态、底层数据缺陷以及组织注意力的分散。企业唯有将AI从孤立的对话框中抽离并嵌入原生流程，依托领域专家与系统集成商打通底层数据，才能实现长期的运营跃升。独立对话框模式的失效与“自信失败”风险 (3:25) Seiberling回顾指出，MXD Process早期利用定制GPT（如融入化工规则的助手“Arlin”）取得了显著成效，将销售线索首联时间从5天压缩至1天，报价速度提升80%，成单率从17%提高到27%。然而，当工具向全员推广时，要求员工切换标签页、登录系统并反复调整提示词的独立对话框架构暴露了严重弊端。团队中仅有约三分之一的技术熟练型员工能够持续获益，其余三分之二则陷入阻碍：部分员工在遭遇劣质回答后迅速放弃；从事电话销售30年的资深人员拒绝使用；而另一部分员工则盲目迷信输出结果，甚至将模型生成的荒谬设备配置直接提供给客户。这种缺乏检验的盲目采纳诱发了Seiberling所称的“自信失败”。例如，一名新任业务经理利用Claude生成了一份排版精美、文笔权威的业绩分析报告，但由于模型忽略了核心业务数据集，其推导结论完全失真。在缺乏人工核验的情况下，该报告直接被分发至全公司，导致管理层一度基于错误信息做出判断。Seiberling强调，AI输出的高权威感极易掩盖实质性逻辑漏洞，给企业运营带来重大隐患。数据管道瓶颈与“搬运数字煤炭”陷阱 (5:45) 为了向管理层呈现关键指标，Seiberling曾亲自通过“氛围编码”（vibe coding）开发了一套高管看板。然而，由于ERP、CRM、货运代理记录及离散电子表格之间缺乏自动化集成，这套工具不仅没有减轻负担，反而迫使Seiberling充当起系统间的人工连接层。每周一早晨，他都必须手动提取、清洗并校验各系统数据，再导入看板之中。他将这种高强度的人工维护戏称为“向AI和看板中铲数字煤炭”。这一经历揭示出中小企业落地AI的底层限制：“没有人能通过升级模型来修复肮脏的数据。”高级大语言模型无法自动克服破碎、混乱的基础数据流，企业所能实现的自动化与智能化上限，本质上完全受制于其底层数据管道的搭建质量。若不优先解决底层接口整合，前端工具只会增加关键人员的操作负担。隐形嵌入工作流与快速落地实战 (8:01) Seiberling主张，推动AI应用落地的关键不是强迫员工学习复杂的提示词工程，而是将模型深度隐蔽在日常业务流程之中。在为包装材料企业Albert Paper Products优化销售流程时，面对依然依赖纸质与Excel的传统环境，团队并未强行推行复杂软件，而是使用会议记录工具完整录制管理层两天的研讨，随后在24小时内自动提炼出目标客户画像与竞争矩阵，使一线销售无需直接面对AI对话框即可享受智能化成果。在MXD Process后续的系统部署中，这一嵌入理念同样展现了极高效率。利用Claude Co-work，团队仅输入联系人表格、公司数据及业务流程文档，在AI提出4至5个针对性问题后，仅耗时4小时便从零完成了一套新HubSpot实例的全面配置，替代了原本需数周的手工映射。同时，基于客户支持数据在1小时内构建的客服看板，成功驱动了预防性支持内容的生成，并将以往每日耗费约2小时的手动数据清洗工作缩减至一键刷新。领域专家主导结合系统集成商架构 (13:11) 针对Heritage Equipment储罐定制报价长达5至6个工作日的瓶颈，MXD Process将工程GPT“Arlin”作为规则引擎，锁定了食品卫生规范与防爆工程等复杂制造约束。在开发模式上，团队通过Claude Co-work快速构建前端，同时联合专业系统集成商Level 12搭建底层架构。这种协同方式仅用3周便跑通了全栈原型，并在历经5周严格的数据校验后正式上线，使耗时数天的报价流程转变为客户可全天候自主调用的即时计算系统。 Seiberling明确反驳了“氛围编码将彻底取代传统SaaS开发”的论调。MXD Process内部仅有一名网络管理员，缺乏专职软件团队，项目的成功依赖于两项支柱：一是让精通业务的领域专家全程在场，确保软件逻辑从第一天起就符合生产实际，同时极大消除内部抵触；二是借助专业系统集成商构筑技术防线。纯粹依靠生成代码搭建的工具在面对复杂企业级交易负荷时极易崩溃，唯有将业务专家与专业工程能力结合，才能产出稳健的生产级应用。中小企业Token经济学与注意力管理 (17:09) Seiberling分析指出，中小企业在考量AI成本时面临着与大型跨国集团截然不同的财务现实。大企业往往需要严格监控巨额的Token消耗与席位费用，但在中小企业的投资回报面前，模型API与软件授权成本基本只是“四舍五入的零头”。中小企业内部最宝贵且稀缺的资产是管理层的精力与一线专家的有效参与时间，一旦核心人员因业务繁重而抽身，AI项目就会陷入停滞。因此，企业的资金投入应重点倾斜于底层数据治理与系统集成服务，而非计较席位支出。基于一整年的实践，Seiberling提炼出中小企业应用AI的五条准则：摆脱对话框将其嵌入原生流程、让领域专家主导场景以压缩时间、认清数据质量决定智能上限、依据组织注意力而非授权费用做预算，以及在全员范围内培养并考核对AI输出的批判性思维与判断力。他总结认为，正如电子商务经历早期阵痛后最终融入常规零售一样，AI只有在不再以突兀的独立工具形态出现、彻底转化为隐形的自动化后台支撑时，才算真正实现业务扎根。 Seiberling认为，中小企业决策者无需被大型机构的复杂叙事所束缚，应依托一线业务常识与批判性判断力推进AI落地。然而，该实践也留下了一个现实问题：当企业高度依赖少数核心管理层的注意力来推动系统集成与流程嵌入时，在缺乏专职技术开发团队的背景下，中小企业应如何建立长效机制以保障系统的可持续演进，仍有待更长时间维度的验证。

Detailed notes — zh

## Background 该视频由 Ai4 频道于 2026 年 8 月 15 日在 YouTube 平台发布。视频内容为工业制造企业 MXD Process 的销售与营销副总裁 Eric Seiberling 的主题演讲，分享了该公司在没有庞大 IT 团队与外部咨询机构支持的情况下，如何将 AI 深度扩展至销售、营销、客户支持以及并购整合等实际业务中的实战经验与教训。 ## TL;DR Eric Seiberling 深入剖析了 MXD Process 从早期基于 Custom GPT 的单点尝试，走向业务流程深度整合与系统化的完整演进历程。他坦诚反思了让员工在聊天框中写 Prompt、数据管道缺失导致“人肉搬砖”以及新手利用 AI 产生“自信的失败”等核心痛点。团队随后转向通过 Claude Co-work、Vibe Coding 并协同领域专家与系统集成商，在数小时内快速搭建 HubSpot 自动化实例，并在 8 周内落地不锈钢储罐在线实时报价系统。最后，他提炼出中小企业落地 AI 的五大铁律，强调 AI 只有彻底走出独立聊天窗口并对用户实现“隐形化流程内嵌”，才能真正扎根发挥价值。 ## Atmosphere & style 整场演讲务实生动、坦率幽默且极具感染力。Eric 毫无传统科技布道者的空洞辞藻，而是以接地气的工业制造一线视角，自嘲“我错了”并直面踩坑泥潭，甚至在现场 PPT 崩溃时脱稿自如调侃。现场叙事节奏明快，既有攻克复杂业务系统时的极速兴奋感，又有对技术泡沫与认知懒惰的敏锐审视，带给听众一种褪去神话外衣、充满清醒务实感的实战共鸣。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=jAO7jheIl0k&t=4s) MXD Process 在第一阶段通过 HubSpot 自动化与化学工程 Custom GPTs 实现了业务突破：线索响应时间从 5 天缩短至 1 天以内，报价效率提升 80%，转化率从 17% 跃升至 27%。 - [02:42](https://www.youtube.com/watch?v=jAO7jheIl0k&t=162s) 依赖独立对话框的 Custom GPT 存在容器缺陷，极度依赖员工自驱力，容易导致试探后放弃、老员工抵触以及盲目迷信 AI 输出三类人员分化问题。 - [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) 警惕 AI 带来的“自信的失败”，缺乏经验的员工可能产出看似专业但底层逻辑全错的报告；同时切勿依赖人工搬砖清洗数据，AI 的上限完全受限于数据管道质量。 - [08:01](https://www.youtube.com/watch?v=jAO7jheIl0k&t=481s) AI 必须走出独立的聊天窗口，无感嵌入到实际业务流程中，靠重构工作流而非强行培训员工写 Prompt 来实现价值。 - [10:51](https://www.youtube.com/watch?v=jAO7jheIl0k&t=651s) 将业务专家与 Claude Co-work 紧密结合，仅用 4 小时便为被收购公司完成整套空白 HubSpot CRM 的业务逻辑搭建，并以 Vibe Coding 快速构建支持工单闭环。 - [13:11](https://www.youtube.com/watch?v=jAO7jheIl0k&t=791s) 结合化工专业模型 Arlen 沉淀工艺规则与前端 Vibe Coding，团队在 8 周内上线不锈钢储罐在线实时报价系统，将 5-6 天的等待周期缩短至秒级。 - [15:20](https://www.youtube.com/watch?v=jAO7jheIl0k&t=920s) 推行 AI 需由深刻理解流程的业务人员主导，并由专业系统集成商把关后端架构，彻底打破传统 IT 部门闭门造车的高摩擦交付模式。 - [17:03](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1023s) 中小企业的 Token 与软件许可成本微不足道，真正稀缺且昂贵的资源是团队的“时间和注意力”，真金白银应优先投向数据治理与系统集成。 - [19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) 中小企业落地 AI 的五大铁律：走出聊天窗口、让专家进入房间、数据决定天花板、预算注意力而非许可证、培养团队的批判性判断力。 - [20:57](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1257s) AI 真正扎根的标志是它“不再看起来像 AI”——如同电子商务演变为日常购物一样，深度融入业务流并对用户隐形的自动化才是最终形态。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=jAO7jheIl0k&t=4s) 工业制造小巨头的 AI 起步与初战告捷 Eric Seiberling 是 MXD Process 的销售与营销副总裁。这是一家年营收约 3000 万美元的工业设备制造公司，主要生产不锈钢储罐与搅拌反应釜——用 Eric 的幽默说法，这可不是“能在路上开、能开炮打仗的那种坦克（Tank）”，而是用于混合化学品、生产各类日常消费品和工业品的大型工业罐体，其应用场景横跨从知名烧烤酱品牌（如 Ray's 烧烤酱及各类超市自营品牌）到航空航天高科技涂层的整个制造链条。三年前 Eric 加入公司时，整个业务流程亟待彻底翻新，他们选择将 AI 作为核心破局工具。在第一年的实践中，团队采取了经典的“爬、走、跑”（crawl, walk, run）渐进策略：从现有工具切入（例如在 HubSpot 中建立自动化流程），并针对行业特性构建 Custom GPTs。因为通用的基础大模型在面对复杂的化学工程问题时表现极差，“就像让一个五年级小学生去解化学题一样糟糕”。这种定制化方案在早期取得了令人瞩目的成效，并在当年的 AI4 大会上引发了强烈反响： - **首触响应周期**：线索响应时间从 5 天大幅缩短至 1 天以内； - **报价生成速度**：效率直接提升了 80%； - **成单转化率（Close Rate）**：从原先的 17% 跃升至 27%。演讲中还出现了一个轻松的小插曲：现场 PPT 投屏一度卡顿崩溃，Eric 笑着调侃自己像在踢足球一样是不是该申请“伤停补时”，随后便在无幻灯片的情况下脱稿继续流畅分享。 ```mermaid flowchart LR accTitle: 第一阶段 AI 改造核心成果 subgraph S1["现有工具自动化"] A["HubSpot 流程自动化"] end subgraph S2["垂直知识定制"] B["化学工程 Custom GPTs （解决通用模型专业缺陷）"] end subgraph S3["关键业务指标突破"] C["响应时间：5天 ➔ 1天"] D["报价效率：+80%"] E["成单转化：17% ➔ 27%"] end A --> B --> S3 ``` --- ### [02:42](https://www.youtube.com/watch?v=jAO7jheIl0k&t=162s) “我错了”：Custom GPT 的容器陷阱与三类员工困境带着第一年“从 0 到 1”的成功经验，Eric 曾自信满满地以为已经找到了中小企业落地 AI 的完美公式，只要照搬推广即可。然而进入第二年后，他给出的核心反思只有三个字：“我错了”（I was wrong）。他将这种落差形容为：刚跑完波士顿马拉松以为冲线夺冠了，抬头一看，眼前竟是一座难度极高的《美国极限体能王》（American Ninja Warrior）障碍赛道。团队遇到的第一个致命硬伤在于：Custom GPT 的工具形态被锁在了“错误的容器”里。传统的聊天机器人（Chatbot）界面需要使用者主动打开网页、登录账号、费尽心思琢磨复杂的 Prompt（提示词），在获得结果后再手动搬运。这种模式完全依赖极高的用户自驱力。团队中只有约三分之一原本就对大语言模型感兴趣的员工用得很好，而其余大部分人在实际使用中迅速分化为三种典型困境： - **试探后放弃型**：培训时觉得很酷，但自己实际输入 Prompt 时由于技巧不足，得不到理想答案，立刻得出“这玩意儿太烂了”的结论并彻底弃用； - **传统抵触型**：有着二三十年工业设备销售经验的老销售员，平日的核心工具就是一部电话，坚信“凭什么让一台电脑来教我怎么卖设备”，自始至终完全无视 AI 工具； - **盲目迷信型**：对 AI 输出缺乏辨别力。公司曾建立了一个名为“Arlin”的化学工程知识库，有员工向其咨询复杂的化学应用方案后将其奉为圣旨，直接向客户兜售；Eric 发现后紧急叫停，因为模型给出的方案要么超出了公司业务范围，要么为了套用公司设备而设计了极其曲折、实际上根本无法交付的方案。 --- ### [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) “自信的失败”与沦为“数字化人肉搬砖工” 随着工具的扩散，另一个最具破坏性的隐患开始浮现——“自信的失败”（Confident Failure）。一位新入职的业务分析经理使用 Claude Co-work 整理公司业务分析报告，生成的图表和文档看似极其惊艳专业，但由于底层缺少关键数据，整个结论从根源上就是错误的。在以往，缺乏底气的新人会主动找上级复核数据；但 AI 赋予了缺乏专业经验者一种虚假的“绝对自信”，导致这份带有致命缺陷的报告未经校验便分发给了全公司，进而误导管理决策。为了解决数据可见性问题，Eric 曾通过“Vibe Coding”快速开发了一套美观实用的管理看板。然而，这并没有实现真正的自动化，反而将他自己变成了系统的最大瓶颈。由于数据分散在 ERP、CRM、物流货运代理（Shipping Broker）系统以及各种零散报表之间，缺乏底层数据打通，Eric 不得不在每个星期一充当“数字化人肉搬砖工”： - 手动从各个异构系统中导出原始数据； - 人工清洗并核对数据准确性； - 手动将数据格式化并上传喂给看板和模型。在这段痛苦的经历后，Eric 总结出了一条至理名言：“没有人能通过升级一个更好的大模型来解决脏数据问题。AI 所能达到的上限，完全受限于你的数据管道（Plumbing）。” --- ### [08:01](https://www.youtube.com/watch?v=jAO7jheIl0k&t=481s) 走出聊天框：从胶带式修补到业务流程再造在公司收购了一家环境服务公司后，新团队急切地希望复制这套 AI 工具。由于当时工作节奏极其紧张、严重缺乏时间，Eric 哪怕心里清楚之前的各种教训，依然下意识地采用了“胶带补丁式”的权宜之计，把旧系统的错误模式在新业务上又重蹈了一遍。这种危机时刻对熟悉工具的路径依赖，让他深刻意识到必须从根本上变革 AI 的交付形态。 Eric 得出的核心破局点是：**必须把 AI 从独立的聊天窗口中解放出来，使其深度内嵌到业务流程中，达到“润物细无声”的赋能状态**。推动 AI 落地靠的绝不是给员工开宣讲会、强行培训写 Prompt，而是彻底重构业务流，将专业知识和算法直接封装在日常工作链条里，让员工只需按照最自然的工作方式行事。为了验证这套“无感嵌入”的流程改造理念，Eric 以兼职首席营收官（Fractional CRO）的身份协助 Albert Paper Products 公司进行销售体系升级。这家制造三级包装材料（如巧克力棒底部的 U 型纸托、高档化妆品盒内的模切内衬）的传统企业，过去完全依赖纸质单据和 Excel 维护销售流程。Eric 的改造实践展示了流程驱动 AI 的惊人速度： - 与客户团队进行了为期两天的深度业务访谈； - 全程使用会议记录 AI 工具（如 Otter、Granola、Fireflies 等）进行完整转录与提炼； - 仅用不到一天时间，便直接自动搭建起涵盖用户画像（Personas）、理想客户画像（ICP）以及竞品分析在内的全套现代化销售体系。 ### [10:51](https://www.youtube.com/watch?v=jAO7jheIl0k&t=651s) 结合领域专家与 Claude Co-work：4 小时上线 CRM 与支持闭环在整合新收购业务的过程中，Eric Seiberling 强调最关键的一点是将业务专家直接召集到同一间会议室。他与全体销售团队坐在一起深入沟通业务细节，一边讨论一边将销售流程进行总结与双向确认，整个梳理过程整整花了三天时间。梳理完成后，面对这家甚至没有部署 CRM 系统的被收购企业，Eric 开启了一次高效的技术实践。他在新买的 Mac Studio 上启动了 Claude Co-work，将三个核心文件上传给 AI：包含客户联系人和公司信息的 Excel 表格，以及刚梳理完成的销售流程文档。Eric 对 Claude 下达指令：“这是一个空白的 HubSpot 实例，请帮我把它搭建起来。”Claude 随后向他提出了四到五个针对性的确认问题。在短短 4 个小时内，一个完全针对该企业业务定制且正常运行的 HubSpot 实例便成功上线。由于整个过程有业务专家全程参与，系统的业务逻辑完全契合实际需求，同时彻底免去了数周繁琐枯燥的底层手动配置工作。 ```mermaid flowchart LR accTitle: Claude Co-work 驱动的业务自动化流程 subgraph S1["数据与流程输入"] A["销售流程文档"] B["客户/公司 Excel 数据"] end subgraph S2["AI 构建核心"] C["Claude Co-work (Mac Studio)"] end subgraph S3["系统与应用落地"] D["HubSpot CRM 实例 (4小时内完成配置)"] E["支持工单分析看板 (Vibe Coding <1小时)"] F["干预措施与解决方案内容"] end A --> C B --> C C --> D D --> E E --> F F -.->|"闭环反馈"| C ``` 随后，Eric 将这一模式复制到了客户支持团队。支持团队的所有工单数据均沉淀在 HubSpot 中，Eric 仅用了不到一个小时的时间通过“Vibe Coding（氛围式编程）”构建了一个专属的数据分析看板。他与一线支持人员坐在一起验证实用性，并对历史工单展开深入分析，自动推导出防止同类问题再次发生的具体干预措施。接着，团队将这些干预措施反馈给 Claude，直接生成用于解决实际问题的支持内容，形成了完整的业务闭环。 - **省去日常低效劳动**：打通底层数据管道后，无需每天耗费两个小时进行数据迁移和清洗，只需点击刷新即可。 - **打通最后一公里**：目前正着手将这些看板与 HTML 报表发布至企业内部的安全门户中，未来计划通过自动化 Agent 实现定时邮件推送与分发。 --- ### [13:11](https://www.youtube.com/watch?v=jAO7jheIl0k&t=791s) 八周交付不锈钢储罐在线报价系统 MXD Process 收购了制造不锈钢储罐的 Heritage Equipment Company。在传统工业制造流程中，客户获取一份定制储罐的报价通常需要等待 5 到 6 天。过去三年里，Eric 一直请求运营团队提供一套哪怕是 Excel 形式的测算工具，但始终未能如愿。这一次，他们借助 AI 打造了一套颠覆性的在线实时报价应用。 ```mermaid flowchart TD accTitle: 不锈钢储罐在线实时报价系统架构 subgraph DataRules["工程约束规则层"] A["化工工程 AI 助手 (Arlen)"] --> B["工艺约束规则 (食品卫生级 / 防爆级等)"] end subgraph DevLayer["协同开发层"] B --> C["前端开发 (Eric) Claude Co-work Vibe Coding"] D["后端与系统架构 (Bruce) Level 12 系统集成商"] end subgraph Output["业务落地"] C <--> D D --> E["3 周完成高可用原型"] E --> F["8 周正式上线 (含5周数据上传与验证)"] F --> G["客户深夜 11 点在线秒级获取报价"] end ``` 这项开发的核心并非单纯套用一个通用的 AI 对话框，而是利用名为 Arlen 的化工工程专业 AI 工具来构建所有底层的制造约束逻辑。例如，用于盛放烧烤酱的食品级储罐与盛放腐蚀性化学品的防爆储罐，在制造工艺、结构与材料约束上截然不同。Eric 利用 Arlen 沉淀出的工程规则，在 Claude Co-work 中通过 Vibe Coding 快速构建前端界面。同时，小型系统集成商 Level 12 的负责人 Bruce 实时同步搭建系统后端。 - **极速敏捷开发**：团队在不到 3 周内就拿出了功能完整、具备高扩展性的原型系统；其余 5 周时间则全部投入到数据导入与工程模型验证中。 - **颠覆客户体验**：系统正式上线仅 9 天就获得了客户与市场的强烈反响。客户无需等待数天，即使在深夜 11 点登录系统输入批次容量与参数，几秒钟内就能实时生成精准的定制报价单。 - **行业反响**：该项目甚至被 Yahoo 等财经媒体报道，打破了传统不锈钢设备制造行业在技术应用上保守落后的刻板印象。 --- ### [15:20](https://www.youtube.com/watch?v=jAO7jheIl0k&t=920s) 业务主导与专家入场：颠覆传统 IT 交付模式 Eric 强调自己是一名在科技行业拥有多年销售和营销背景的“业务人员”，而非专业 IT 人士。MXD Process 内部没有任何庞大的 IT 资源，仅有一名负责维护网络的基础人员。这种非传统背景反而让他摸索出一套高效的落地路径。在这一过程中，让业务专家与专业系统集成商（SI）全程在场的价值体现在两个核心维度： 1. **从一开始就用正确的方式构建**：避免由于技术人员缺乏业务理解而造出错误系统。 2. **保障落地采用率**：一线人员全程参与了系统的设计与搭建，系统上线后会立刻被积极拥抱，而不是遭到排斥。传统 IT 部门往往习惯于“收集所有业务需求，然后钻进小黑屋闭门造车，几个月后出来喊一声‘看，大功告成’”，这种模式往往伴随着高昂的沟通摩擦与失败率。与此同时，Eric 提醒道，仅凭业务人员自己的 Vibe Coding 是不够的，必须有像 Level 12 这样的专业系统集成商把关后端架构与用户体验，确保原型转化为真正健壮的企业级生产应用。 --- ### [17:03](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1023s) 中小企业 Token 经济学真相：最大的成本是“时间和注意力” 谈及 AI 落地预算与所谓的“Token 经济学（Tokenomics）”，Eric 结合自己在亚马逊（Amazon）和宝洁（Procter & Gamble）等大型企业的工作经历进行了对比。大型跨国企业确实需要极其严密地核算软件许可证、Token 消耗以及基础设施成本，但对于广大中小企业（SMB）而言，如果使用的是基础工具链，这些 Token 与软件成本“完全只是四舍五入的零头（Rounding errors）”。在中小企业中，真正昂贵且需要严格做预算的资源是**团队的时间与注意力**。 > “注意力之所以被称为‘付出注意力（Paying attention）’，是有原因的。如果我因为日常事务过于繁忙而搁置了 AI 项目，整个企业的 AI 进程就会立刻停滞；如果无法把业务专家抽调到会议室，大家总是在喊‘太忙了’，企业就永远无法实现改进。” 中小企业真正需要投入真金白银预算的地方只有两个： - **数据清洗与治理**：如果数据基础是一团乱麻，AI 产生的结果将毫无意义甚至误导决策。 - **专业系统集成支持**：针对市场上声称“在 SaaS 灭亡时代，任何人都能靠 Vibe Coding 搓出一个 CRM”的论调，Eric 直言这纯属无稽之谈，缺乏架构设计的脆弱代码在实际业务压力下瞬间就会崩溃。 --- ### [19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) 中小企业落地 AI 的五大核心法则 Eric 总结了这一年带领团队在真实业务中踩坑与突破后沉淀下来的五条铁律： | 核心法则 | 核心主张与实操要点 | |:--- |:--- | | **1. 走出聊天窗口** *(Get it out of the chat window)* | 仅仅把 AI 停留在对话框里需要员工主动保持使用意愿，一旦业务繁忙，人们就会迅速退回习惯的旧工作流。必须将 AI 能力直接内嵌到业务流程与系统中。 | | **2. 让专家进入房间** *(Get an expert in the room)* | “AI 能够压缩专家的时间，但无法凭空创造专业知识。”必须由具备深厚行业经验的专家主导业务逻辑的梳理。 | | **3. 数据决定天花板** *(Your ceiling is your data)* | 没有扎实的数据管道和干净的基础数据，任何 AI 实践都会溃败，甚至会输出错误的业务判断导致企业蒙受财务损失。 | | **4. 预算注意力，而非许可证** *(Budget attention, not licenses)* | 不要过度纠结工具席位费。Eric 曾审查团队的 ChatGPT 使用率并果断削减了 10 个闲置账号；只要核心流程跑通，工具本身的账单无关紧要。 | | **5. 培养批判性判断力** *(Teach judgment)* | 谨防 AI 带来的认知懒惰与思考能力退化。企业在招聘、考核乃至淘汰员工时，必须考察其是否具备探究“结果是否正确、为什么正确、是否可信”的审慎判断力。 | ### [20:57](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1257s) 让 AI 真正落地的核心：隐形化与流程集成 AI 真正能够扎根并持续发挥价值的关键，在于它“不再看起来像 AI”（"AI only sticks when it stops looking like AI"）。回顾以往经历的互联网与互联网泡沫（Dot-com boom）等重大技术变革，商业模式随之发生了翻天覆地的变化。以电子商务为例，在宝洁（Procter & Gamble）工作期间，行业内曾一度弥漫着“实体零售即将消亡”的论调，但事实证明实体零售并没有消亡，而是演变成了全新的形态；如今人们进行线上购物时，早已不再觉得这是某种刻意为之的“电子商务”，它已经自然而然地变成了“我们日常购物的方式”。同样的逻辑也完全适用于当今企业运用 AI 赋能业务的场景。企业如果想在 AI 转型中胜出，就必须把 AI 从单纯独立的“聊天机器人（chatbot）”外壳中剥离出来，将其深度集成并无缝嵌入到实际的业务流程中。当它对终端用户来说完全隐形、变得如行云流水般“自动化（automatic）”时，才是企业真正获得胜利的时刻。 ```mermaid flowchart LR accTitle: AI 落地演进路径 subgraph S1["初期形态"] A["独立 Chatbot"] --> B["用户感知明显 / 割裂"] end subgraph S2["终极落地形态"] C["嵌入业务流程"] --> D["对用户隐形且自动化"] end S1 --> S2 ``` ### [21:44](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1304s) 演讲总结、交流互动与最后的建议随着在 MXD Process 一年以来的 AI 实践与规模化推广告一段落，演讲者希望明年能再次回到大会舞台，分享更多后续的发展与经验。在留出联系方式时，演讲者幽默地向台下听众划定了沟通规则：欢迎非竞争对手通过 LinkedIn 建立联系并探讨交流；如果是竞争对手找来，“我依然会和你聊，只是我会告诉你一堆全都是错的信息”；而如果是向他推销产品的供应商，则只有 30 秒的时间来进行快速推介。在演讲的尾声，演讲者给所有试图在中小企业中推行 AI 的管理者和从业者留下了最核心的建议：相信自己，相信自己的商业判断力，充分吸收并运用这些在实战中摸索出来的经验教训，任何团队都能够复制同样的成功。 - **历史技术变革的借鉴**：如同电子商务最终褪去概念外衣成为日常购物习惯一样，AI 也必须融入日常业务运转。 - **业务集成标准**：脱离单纯的聊天界面，做到对操作者隐形，实现真正的自动化交付。 - **给从业者的寄语**：在充满不确定性的技术浪潮中，要“相信自己，相信自己的判断”（"trust yourself, trust your judgment"）。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=jAO7jheIl0k&t=4s) 就像让一个五年级小学生去解化学题一样糟糕 — Eric Seiberling > [02:42](https://www.youtube.com/watch?v=jAO7jheIl0k&t=162s) 刚跑完波士顿马拉松以为冲线夺冠了，抬头一看，眼前竟是一座难度极高的《美国极限体能王》障碍赛道 — Eric Seiberling > [02:42](https://www.youtube.com/watch?v=jAO7jheIl0k&t=162s) 凭什么让一台电脑来教我怎么卖设备 — Eric Seiberling > [05:45](https://www.youtube.com/watch?v=jAO7jheIl0k&t=345s) 没有人能通过升级一个更好的大模型来解决脏数据问题。AI 所能达到的上限，完全受限于你的数据管道（Plumbing） — Eric Seiberling > [08:01](https://www.youtube.com/watch?v=jAO7jheIl0k&t=481s) 必须把 AI 从独立的聊天窗口中解放出来，使其深度内嵌到业务流程中，达到“润物细无声”的赋能状态 — Eric Seiberling > [10:51](https://www.youtube.com/watch?v=jAO7jheIl0k&t=651s) 这是一个空白的 HubSpot 实例，请帮我把它搭建起来 — Eric Seiberling > [15:20](https://www.youtube.com/watch?v=jAO7jheIl0k&t=920s) 收集所有业务需求，然后钻进小黑屋闭门造车，几个月后出来喊一声‘看，大功告成’ — Eric Seiberling > [17:03](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1023s) 注意力之所以被称为‘付出注意力（Paying attention）’，是有原因的。如果我因为日常事务过于繁忙而搁置了 AI 项目，整个企业的 AI 进程就会立刻停滞；如果无法把业务专家抽调到会议室，大家总是在喊‘太忙了’，企业就永远无法实现改进 — Eric Seiberling > [19:00](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1140s) AI 能够压缩专家的时间，但无法凭空创造专业知识 — Eric Seiberling > [20:57](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1257s) AI only sticks when it stops looking like AI（AI 只有在不再看起来像 AI 时，才能真正扎下根来） — Eric Seiberling > [21:44](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1304s) 我依然会和你聊，只是我会告诉你一堆全都是错的信息 — Eric Seiberling > [21:44](https://www.youtube.com/watch?v=jAO7jheIl0k&t=1304s) trust yourself, trust your judgment（相信自己，相信自己的判断） — Eric Seiberling ## Mentioned Eric Seiberling, Bruce, MXD Process, HubSpot, Custom GPTs, Claude Co-work, Arlin / Arlen, Albert Paper Products, Otter, Granola, Fireflies, Heritage Equipment Company, Level 12, Yahoo, Amazon (亚马逊), Procter & Gamble (宝洁), ChatGPT, LinkedIn, AI4 ## Video description Eric Seiberling - VP of Sales and Marketing, MXD Process Last year, I showed how we started using AI at a small industrial company—with no IT team and no fluff. Since then, we’ve scaled those early wins in sales, marketing, support across the company and two acquisitions — without consultants or custom software. But it wasn’t easy. In this talk, I’ll share what we broke, what we fixed, and the messy, behind-the-scenes lessons we learned trying to make AI stick in a real business. You’ll walk away with a practical playbook for growing AI inside a small team—what to automate, where humans still matter, and how to budget for the things no one warns you about. [Source](https://www.youtube.com/watch?v=jAO7jheIl0k)

Source quotes

[0:04] Eric Seiberling: 就像让一个五年级小学生去解化学题一样糟糕 [2:42] Eric Seiberling: 刚跑完波士顿马拉松以为冲线夺冠了，抬头一看，眼前竟是一座难度极高的《美国极限体能王》障碍赛道 [2:42] Eric Seiberling: 凭什么让一台电脑来教我怎么卖设备 [5:45] Eric Seiberling: 没有人能通过升级一个更好的大模型来解决脏数据问题。AI 所能达到的上限，完全受限于你的数据管道（Plumbing） [8:01] Eric Seiberling: 必须把 AI 从独立的聊天窗口中解放出来，使其深度内嵌到业务流程中，达到“润物细无声”的赋能状态 [10:51] Eric Seiberling: 这是一个空白的 HubSpot 实例，请帮我把它搭建起来 [15:20] Eric Seiberling: 收集所有业务需求，然后钻进小黑屋闭门造车，几个月后出来喊一声‘看，大功告成’ [17:03] Eric Seiberling: 注意力之所以被称为‘付出注意力（Paying attention）’，是有原因的。如果我因为日常事务过于繁忙而搁置了 AI 项目，整个企业的 AI 进程就会立刻停滞；如果无法把业务专家抽调到会议室，大家总是在喊‘太忙了’，企业就永远无法实现改进 [19:00] Eric Seiberling: AI 能够压缩专家的时间，但无法凭空创造专业知识 [20:57] Eric Seiberling: AI only sticks when it stops looking like AI（AI 只有在不再看起来像 AI 时，才能真正扎下根来） [21:44] Eric Seiberling: 我依然会和你聊，只是我会告诉你一堆全都是错的信息 [21:44] Eric Seiberling: trust yourself, trust your judgment（相信自己，相信自己的判断）

Source quotes

[1:29] Eric Seiberling: having a fifth grader try to do chemistry. [1:29] Eric Seiberling: I was wrong. [3:25] Eric Seiberling: Wait, wait, wait, wait, wait, we don't sell that. Stop. [5:45] Junior Employee: Hey Eric, could you check my work? [5:45] Eric Seiberling: shoving digital coal into the AI and into the dashboard. [5:45] Eric Seiberling: Nobody's ever fixed dirty data by upgrading to a better model. [10:51] Eric Seiberling: This is a blank HubSpot instance. Please stand it up, [19:00] Eric Seiberling: AI compresses experts' time, but it doesn't manufacture expertise. [20:57] Eric Seiberling: AI only sticks when it stops looking like AI. [21:44] Eric Seiberling: Trust yourself. Trust your judgment. Take these lessons and you'll be able to do the same things that we've been able to do.

Original transcript

[0:04] My name is Eric Cyberlang. I'm the vice president of sales and marketing at MXD Process. Now, [0:08] you probably have never heard of us. We're just a $30 million industrial equipment company. And so, [0:14] we make stainless steel tanks, not the ones that you drive around and shoot things, [0:18] but the big vats that you end up mixing up chemicals in order to be able to make the products [0:21] that you know and love. So everything from barbecue sauce all the way to aerospace coatings, [0:26] we end up making the equipment that goes into the factory in order to be able to do that. [0:30] So um I was here last year and I ended up giving a presentation on our beginning of our journey of [0:36] AI. So I got there now last year it was two years now three years and the business needed overhaul [0:45] and so trying to get everything fixed and we ended up using AI to do it and so [0:51] I told that story last year. This year is the continuation of that story. [0:55] Barbecue, we actually make the tanks that then can be used by all sorts of barbecue sauce companies. [1:01] So you talk about Rays um Ray's barbecue sauce. We ended up doing also for a number of um uh [1:06] basically it's um grocery stores end up having private label and just any number of those. So [1:14] Oh wow. Okay. Feel like in soccer, do I get extra time? It flopped. Yeah, my PPT flopped. I'm sorry. [1:24] But yeah, so Oh, here I'll just keep going. So, ended up, anyway, I ended up talking about our [1:29] journey last year. We ended up taking an approach of crawl, walk, run. Everybody gets to hear that, [1:33] right? But it's we talked about how we ended up starting with things that we already had. So, [1:39] how did I do automations in HubSpot? We started creating custom GPTs because chemical engineering [1:45] could not be handled by GPT3. It's like having a fifth grader try to do chemistry. It was bad. So, [1:51] we end up making custom GPTs and we were able to achieve really great results. And one of the [1:57] results, you know, we ended up going from a lead time from 5 days down to one first [2:01] contact. We ended up being able to improve our quoting speed 80%. And the biggest one of all is [2:07] we were able to improve our close rates from 17 to 27%. Now, for a small company, is that a big deal? [2:14] It is. It absolutely was. and it's been a gamecher for our business. And so finished up my talk, [2:20] everyone's like, "Ah, that's amazing. Can I call you? Can I meet with you?" All these things. [2:24] And I'm like, "Yeah, great. We think we've got it figured out. We went from zero to one." [2:29] And it's like, "Awesome. We've got this figured out. Now we can make this grow." And so [2:36] patting myself on the back, leaving AI4, and then I'm saying, "We've got the right strategy. Great. [2:42] Now we can move forward." Well, I have one thing. I have three words to tell you for our [2:47] second year of our journey. I was wrong. Let's see if I can show you the slide of I was wrong. [2:56] So these were the So okay, go back one. I just want people to see the results. [3:01] So those were the results that we had. And for anybody that runs a small business, you know, [3:06] that was pretty amazing. But again, like I said, I was wrong. It's like I just got done running the [3:12] Boston Marathon and then I end up going and now I've got this amazing OP, you know, American Ninja [3:17] uh obstacle course to try to run. So, this talk is about what we've learned through that. [3:25] The first thing that we ended up learning is we had these great custom GPTs and it was an [3:29] amazing great tool. The problem was it was in the wrong container as we started to deploy it and all [3:35] those things. Think about what what a what does a chatbot require you to do? go onto a website, [3:41] log in, type your questions, prompt it the right way, which is a whole trick in and of itself, [3:46] be able to get the answer, and then go. It's great, but it was kind of locked behind. And so, [3:51] what does it require? It requires a motivated user in order to be able to go in and actually [3:55] do the work. Now, there were people in our team that were already familiar with LLMs. [3:59] They really liked it. They trusted it. And so, boom, they went and they went ran off. And it [4:03] was about a third of our team did really well with it. But what about the others? [4:11] The whole thing is that people have to be motivated in order to use chat bots. [4:15] And so that was our first issue of we had this great knowledge locked in a box [4:19] that people had to go and unlock every single time. So our users fell in one of three places. [4:26] So number one, we trained everybody how to do it. They wow that's cool. So they went and they [4:30] prompted it and all that stuff. They didn't get the same results and they're like this sucks, [4:34] right? And so they like boom, I'm done. The second group were people that we had sellers that have [4:40] been selling industrial equipment and tanks for 30 years. Their tools really are it was a phone. [4:47] And so they're like, you know, some computer is going to tell me how to sell industrial equipment. [4:52] So they never they just ghosted it. They never tried. And the third one, which is even crazier, [4:56] was, you know, chatbots are not perfect. And so what ended up happening was is when someone [5:02] used we had a kind of it's a knowledge chemical engineer knowledge base which was called Arlin is [5:08] what we called it. And so with Arlin they type in and say how do I end up doing this? And it [5:14] was a relatively complex chemical thing. And so they took it as gospel and they started to sell [5:19] it to the customer. And I'm like wait wait wait wait wait we don't sell that. Stop. [5:25] or it was such a convoluted roundabout solution to use our equipment because that [5:28] was one of the guardrails we had. It would be impossible for us to actually deliver. And so [5:34] those were the three ways and that's what you have to work through in the hands of a skilled user [5:37] who's motivated to do it. Custom GPTs work amazing but beyond that you can get yourself into trouble. [5:45] Now here's the biggest one that a lot of us may or may not have had to deal with. [5:54] This is the issue that scared me the most. So, we had a brand new manager that ended up coming [5:57] to work for me. He was really excited doing business analysis, all those things. So, I used [6:02] Claude Co-work in order to be able to analyze the business and be able to pull it all together and [6:05] built an amazing report. And it would have been great, but it was missing data. So, it was wrong. [6:14] Now, in the hands of someone that's unsure, what happens? Hey, Eric, could you check my work? [6:19] Is this right? Are you sure this is good? all that stuff. But the problem is AI allows [6:27] I won't say people that don't have the knowledge to feel like they have very confident answers [6:31] and so they start to distribute it. There's no check and when it's a manager it gets [6:37] distributed across the company and then the company's starting to act on that. [6:41] So that's one of the big dangers that we ended up having was really that. And so again, [6:48] confident failure is something you have to watch out for as you get these tools deployed. [6:54] So I ended up vibe coding him a dashboard and it was a cool dashboard and it worked and it was [6:59] really beautiful. And oh by the way, it was really good to make decisions. But here was the problem. [7:04] Instead of having the AI be able to do the work, guess who was doing the work? [7:08] Me. I was the guy shoving digital coal into the AI and into the dashboard to make sure [7:14] that it worked because I had to extract the data because it was multissource across our ERP, CRM, [7:19] our broker, our shipping broker information and um some other reports. Had to bring it [7:25] all in and then I had to make sure it was right. And so every single Monday I had to go through [7:30] pull the report, upload it. And so guess who was the integration layer? That was me. [7:35] And so that's great if I have a lot of time, but in my role, I really don't. [7:41] Nobody's ever fixed dirty data by upgrading to a better model. [7:46] Your upper limit of what you can do in an AI is limited by your plumbing. [7:51] And that's the whole thing. If you don't have clean data and you don't have good plumbing in [7:54] order to get that data to your model, you're in trouble. And so, guess what? I was in trouble. [8:01] So, here's the other thing, and I'm kind of embarrassed to say this. So, we acquired [8:04] a uh industrial equip or an environmental services company and they're like, "Hey, [8:09] these tools these tools you have are really great. Can we have one?" And I knew all of these lessons. [8:16] I knew them all. But I ended up do what I did the duct tape solution. I had no time. They really [8:22] needed something. So, what did I do? I repeated the exact same mistake I did before. Why? Because [8:27] I'm dumb? No. Because I had no time. And so when life is in crisis and all those things, what do [8:33] we do? We grab for what we know. We say, "Oh, this will be good for now and then we don't have much [8:39] time and attention to be able to go back and fix it later." And so we can end up creating problems. [8:47] Where do we go from here? So that was kind of the whole situation of what I needed to deal with. [8:52] And so there's really one big lesson from our standpoint. It could be different for everybody [8:56] else, but you have to get AI out of the chat window and make it so it actually empowers [9:03] what you do without seeing it. The way you drive adoption isn't by convincing people and motivating [9:10] people in order to be able to to, you know, use this tool. It's great. All those things, [9:15] take the time, learn it. No, you're better off really reinventing your business processes. So [9:20] then what ends up happening is is the knowledge is embedded in and all they need to do is their job. [9:25] And I think that is something that we all have to think about as we're entering in this AI space. [9:30] And I know vendors are promising the world of all this will do. But it's the same thing that we saw [9:34] with chat bots. They don't know our business. I'm in a um long scale. It's a long lead capex [9:42] capital intensive process that requires a lot of custom engineering and a lot of custom knowledge [9:47] in chemistry. And there's really not many tools out there that can do that for me. Now [9:52] some people be like, I can do it. I haven't seen it yet. Okay. And so, let me talk to you. Now, [9:57] this was kind of one of the lessons. And so, what did we do? So, I'm also the fractional CRO for [10:03] other companies. And so, Albert Paper Products is one of them. They do tertiary packaging. So, [10:08] when you open up a candy bar, you know, the little the little U board that's in there that keeps the [10:12] chocolate kind of in there. That's some of the stuff that they make. Or when you get a cosmetics [10:16] and you open up the box and it has this little die cut, that's what they do. [10:21] So, but they were in the Excel and paper part of running their their sales process. And so, [10:28] I was asked to go and how do you fix it? [10:32] So, here's what we did. I ended up meeting with them for two days and I had Otter on the [10:36] whole time or granola, Fireflies, pick your favorite, right? But then I ended [10:40] up being able to summarize it and was able to build their the sales process in under a day. [10:46] I had personas, I had ICPs, I had competitive analysis, I had everything all built out. [10:51] And so, you know, with the experts in the room, and that's the key part. I had all [10:56] the sales people there. We all talked about it. As I started to summarize, [10:59] I shared it back with them. Made this really great give and take and it was literally three days. [11:05] So, then I had this document. Now, what do I do with that document? Well, [11:08] they had no HubSpot. So, I ended up starting to stand up their HubSpot. But here was the fun part. [11:12] I ended up firing up Claude Co-work on a new Mac Studio that I bought and I ended up uploading they [11:18] had their contacts and companies and an Excel and then I ended up having their sales process. [11:22] I uploaded all three and they said, "This is a blank HubSpot instance. Please stand it up." [11:27] It asked me about four or five questions and over four hours I had a fully running HubSpot instance [11:33] that was configured for their business. I had the experts in the room and I built it the way that [11:37] they wanted it, but it got me out of all the grunt work of actually setting it up. saved me weeks. [11:44] Second one, we all remember the dashboard instance I had before. So I ended up doing the exact same [11:50] thing for our support team. So all of our support data is in HubSpot. That's what we end up using. [11:56] But I ended up vibe coding it in less than an hour. But the advantage was and I worked [12:01] with the support person. We say does this work? And if you see the small box there, [12:05] I know it's hard to read, but I actually did an analysis of all of our tickets and say [12:09] this is the way these are the interventions you need to prevent these things in the future. [12:14] And now we take those and we put them back into Claude and we end up actually building the content [12:19] that will actually solve the problem. So we close that loop and then we can go back and measure and [12:23] say does this continue? So the difference was we did the same process, sat down with the owner, [12:30] worked through the whole thing, did all that, but the plumbing was there. And so because of that, [12:37] I don't have to do anything. I don't have to spend two hours a day migrating data, cleaning it up, [12:42] doing all those things. All I have to do is hit refresh. Now it's in co-work. Okay. So [12:46] the last mile is how do I publish this inside a a corporate secure portal so that way it's you know, [12:53] I don't even have to distribute it. you know, the HTML file to everybody once a week. I know I can [12:57] set up an agent to do that to do the email or do other sorts of things and I'm getting to that. Um, [13:03] but it's really been a difference of doing this work very differently. [13:11] Here's the one that's most fun and I'm really proud of. Now, if you watch the screen here, [13:15] this is an application we vibe coded in 8 weeks. What it does is it allows you to quote a stainless [13:21] steel tank. So you can end up putting in the batch size. You can end up doing all [13:25] those things because we acquired a company called Heritage Equipment Company. They make stainless [13:29] steel tanks. So quotes normally take five or six days in order to get it, bring it back and all [13:35] those things. Okay. I asked operations for three years to help me build something. Give me an Excel [13:40] spreadsheet so I can quote it. Now this is the way AI played into this. This is not an AIdriven tool. [13:47] What we use is Arlin. remember that chemical that the chemical engineering kind of chat tpt we had [13:53] ended up using it to be able to build all the constraints for us to build the tank. [13:58] So if something is sanitary because it's barbecue sauce there are specific rules [14:02] that you have to build the tank in certain ways you have to build it. If it's explosion proof [14:07] because it's costic chemicals or all those things you have to build it a different way. [14:11] And so we used Arlin in order to develop those rules that then we vibe coded in uh Claude Co-work [14:19] in order to create the front end. But here's the difference. The difference was we had [14:23] someone from level 12. They're a very small system integrator as our owner and I am vibe coding this. [14:33] The Bruce who's the owner of Lever 12 was sitting there building the back end at the same time. [14:38] So we had a full working app prototype in less than three weeks and it was scalable to be [14:43] able to do that. The other five weeks was spent uploading data and validating the model. So we [14:49] just launched this 9 days ago and so the response from the company from our customers have been very [14:55] strong because normally they have to wait days for quotes. Now they can go on at 11 o'clock at night, [15:00] fill it out and get a quote. Literally you will watch it pop right up as you see it roll through. [15:04] So, sorry I have it on repeat because I know I talk longer than it would take to do the quote. [15:08] If you want to, that's the that's the QR code to actually try it yourself. If you do and you're not [15:13] in the chemical business, just put after your name um AI4 so that way my sales people don't bug you. [15:20] Okay? So, but again, the key learning really is you got to get the experts in the room, [15:28] right? It's one of those things that and and I am a business guy. I am not an IT guy. [15:35] We don't have any IT resources. I got one guy that manages our network. That's it. [15:38] I'm a business guy. Okay. And so I've kind of just ground through it. I knew [15:42] enough about tech and I've worked in the tech industry for years as a marketer and as sales. [15:48] So I had enough background to feel like why not? Let's go do this. But you have to have the experts [15:53] in the room for two reasons. Number one, you build it the right way. Number two, they accept [16:00] it coming out the gate because they help build it. And a lot of times it T organizations want to give [16:06] me all your requirements. We'll go through, we'll go in our little go in our little closet, we'll [16:10] build it, and then we'll bring it back out and go, tada. Or they'll lend now, let's work through it, [16:15] iterate, right? You're better off doing the whole process with somebody that's in the room. [16:21] And so that's what was able to be much more successful. The other pieces in all this [16:27] is that the other expert in the room is having somebody that can actually turn something that's [16:31] vibe coded into an actual real application. And that's what Level 12 did for us. And they were [16:36] very inexpensive to work with. We have a strong relationship with them. And if you guys want to [16:40] meet them, just let me know and I'll I'll make the connection. But in eight weeks and they learned a [16:45] lot about what we do chemical engineering wise and they made sure we didn't do anything stupid from a [16:49] technical or UX standpoint. Okay. So that's really important and that's why we've shipped it and [16:55] we've we've been picked up by Yahoo and a couple other things just in the work that we've done [16:59] because stainless steel equipment manufacturers are not known to being on the cutting edge of [17:03] technology. So but that also comes to this other piece if you want the experts you have to make [17:09] sure your data is right all those kinds of things you know this always comes back to what thing that [17:14] everybody brings up budget or tokconomics as people are calling it for an SMB only. Now this [17:22] is very different. I worked at Amazon. I worked at um Proctor and Gamble do systems and I know [17:28] they had big token they have to really worry about what are software costs what are all those things [17:34] for an SMB for the most part in particular if you're using a lot of the basic tools. [17:38] I'm going to say this and people can disagree with me but those are rounding errors. [17:44] You know, if I talk about software licenses, if we talk about tokens, if we talk about those, [17:47] I'm not going to spend that much. And if I just need to make sure I get the ROI, the biggest [17:52] thing, the biggest cost, the biggest thing you have to budget for is time and attention. [17:56] It's called paying attention for a reason. If you don't, for example, if I get too busy and [18:01] I'm supposed to be working on this thing, guess what happens to our whole AI initiative? It stops. [18:06] If I can't get the experts in the room and time goes, we're too busy, we're too busy, then I can't [18:10] improve. I can't get better. And so you really have to think about that and really end up going [18:17] through and say how do I invest my people's time? How do I ensure that I have enough attention? [18:24] Now the place you may have to spend budget which I would highly agree with is on data cleanup [18:29] because if you don't pay attention to that, your AI is never going to work. [18:34] So you have to have the time to experiment. You've got to have the time to build because [18:38] then and you also need sometimes to have the SI in the room because that's how you build [18:42] something that survives the test of time because anybody who tells me in the SAS apocalypse and [18:47] oh I can vibe code my own CRM doesn't know what they're talking about. I'm sorry. Those things [18:51] collapse under any kind of weight and pressure. This is how you build things that matter. So [19:00] here are five lessons that I ended up learning through this and it's just I think a really good [19:04] takeaway for this talk. My screen disappeared. No. Number one, get it out of the chat window. [19:09] Okay, just like I said before, that was the biggest issue for us is if it's [19:13] locked up in the chat window, people have to be motivated to do it. And when people get busy, [19:17] they go back to what they know. Build it and integrate it into your work process and change [19:23] your work process because of what AI does for you. Now, you still need the human in the loop, [19:28] but get it out of the chat window. Number two, get an expert in the room. [19:32] AI compresses experts time, but it doesn't manufacture expertise. [19:39] Okay? And so you really have to sit there and think about that and just do that. Well, [19:43] third is your ceiling is your data. If you don't have good plumbing, if you don't have clean data, [19:49] anything you try to do from an AI perspective is going to bomb or worse, it's going to have you [19:54] make bad decisions that's going to cost you money. And so, take the time to do that. Fourth, budget [19:59] attention, not licenses. Now, we've gone back through and I've looked at my chat GPT um kind [20:05] of group and who all using because I looked at usage rates. I just took 10 people out of our pool [20:10] because they're not really using it. And guess what? If I build it into my process, I don't care, [20:16] right? Fifth is really teach judgment. And I think this is the thing. This is the big [20:21] danger with AI. You know, we talk about loss of cognitive load and you know, or people being able [20:26] to think through problems. This is what you have to, you know, you've got to hire for it. [20:31] You've got to evaluate people on it. And sometimes you have to fire people for it [20:34] if they're not curious and if they're not willing to go through and take the time to say, [20:38] "Is this right? Why is it right? Do I believe it?" And really get that judgment built in themselves [20:44] to know when they get a result, hey, does this make sense? [20:48] You're going to run yourself into problems. So, those are kind of the big five that I [20:52] that I kind of thought through for this year. And hopefully next year we'll talk about five more. [20:57] Here's the thing. AI only sticks when it stops looking like AI. Okay? I mean, I live through the [21:05] internet, the.com boom. I live through a number of changes. When you talk about how we do business, [21:10] e-commerce, I mean, I worked at Proctor and Gamble and they're like, "Retail's going to die. [21:15] Is retail dead?" No, it's just changed, right? But it doesn't feel like e-commerce and, you know, [21:22] you know, all these other things. It just feels like the way that we shop. Okay? And [21:26] so the same thing goes for the way that we run our businesses. And so when you can get it out of the [21:32] out of a chatbot, integrated into your process, and it basically is invisible to the user, it's [21:37] it's automatic as I like to call it. That's when you're going to win. So that's kind of our story. [21:44] And so stay tuned next year. Hopefully I'll be able to come back and share that. [21:48] But if you want to get a hold of me, if you'd like to talk about and you're not a competitor, [21:52] there's my LinkedIn. If you are a competitor, I'll still talk to you. I'll just give you all [21:56] the wrong information. And if you're a vendor, you'll have 30 seconds to pitch me. That's it. [22:01] Okay? But for the most part, it's trust yourself. Trust your judgment. Take these [22:07] lessons and you'll be able to do the same things that we've been able to do. So, thank you.